Chinese AI models in 2026: DeepSeek, Qwen, Kimi and GLM, and are they safe?

Part of our AI models guide: 11 posts on choosing and running AI models.
Is it safe to use a Chinese AI model? It depends where it runs: a lab's own app or API sends your prompts to that company, while open weights on GPUs you control keep them on your hardware. The question matters because most of the strongest models you can download are now Chinese, led by DeepSeek, Alibaba (Qwen), Moonshot AI (Kimi) and Z.ai (GLM), and eight of the ten models developers used most on OpenRouter in mid-November 2026 came from Chinese labs. The best, Kimi K3.1 Pro, scores 46 against 65 for the top closed US model on Artificial Analysis's Intelligence Index v5, at about a seventh of the cost per task. US government tests found some of these models easier to jailbreak and more likely to echo Chinese Communist Party narratives, so test before production. Restrictions so far cover the labs' apps and government work, including US federal contractors, not other private use, and most licences allow commercial use.
Weights are the numbers that make up a model; open-weight means you can download them, which is not the same as open source (see open-weight vs closed models). Most of these models use a small active share of their total parameters for each token (the smallest unit of text an AI model reads and writes, often a short word or part of one). Context is how much text a model reads at once.
Chinese LLMs at a glance
Lab | Latest open weights (2026 unless stated) | Total / active parameters | Licence |
|---|---|---|---|
DeepSeek | V4-Pro-0813 (13 Aug); V4.1-Flash-1017 (17 Oct) | 1.6T to 1.7T / 49B; 552B / 8B to 16B | MIT |
Alibaba (Qwen) | Qwen3.8 2.4T (Aug); Qwen3.8-27B (5 Aug); Qwen3.8-Flash-Next (24 Aug) | 2.4T / 95B; 27.8B, all active; 125B / 6B | Custom; Apache 2.0; Qwen Community License |
Moonshot AI (Kimi) | Kimi K3.1 Pro (early Nov) | 2.8T / 104B | Custom |
Z.ai (GLM) | GLM-5.3 (25 Aug); GLM-5.3-Flash | About 753B (active not stated); 320B / 18B | Custom, MIT-based; MIT |
MiniMax | M3.1 Flash (17 Oct); M3 (2 Jun) | About 428B / 23B each | MiniMax Community License |
Tencent (Hy) | Hy4 flash (17 Nov); Hy4-preview (28 Aug) | 220B / about 12B; 770B / 49B | Tencent Hy License; Apache 2.0 |
StepFun | Step 5 Preview (mid-Oct) | 600B / 27B | StepFun Model License |
Xiaomi (MiMo) | MiMo-V2.6-Pro (21 Sep); V2.6-Flash | 1.02T / 42B; 309B / 15B | MIT |
Baidu (ERNIE) | ERNIE-5.5-Flash-Base (about 5 Nov) | 280B / 14B | ERNIE licence (no rights in the EU, UK or South Korea) |
ByteDance (Seed) | Seed-OSS-36B (Aug 2025) | 36B, all active | Apache 2.0 |
Meituan (LongCat) | LongCat-2.5 (4 Nov) | 1.6T / about 48B | MIT |
Ant (Ling) | Ling-3.1-flash (13 Oct) | 560B / 25B | MIT |
Most read up to 1M tokens of context. ERNIE-5.5-Flash-Base is a base model with no chat tuning.
Open by default, but not always: Alibaba (Qwen4-Max), Baidu (ERNIE 5.1) and ByteDance (Seed 2.1 Pro) keep their newest flagships closed for now. Tencent promises Hy4's weights and Alibaba smaller open Qwen4 models, neither with a date.
Is it safe? Where your data goes
A lab's own app or API. DeepSeek's privacy policy (10 February 2026) says it stores personal data in China, may share it with law enforcement and public authorities, and trains on prompts and files unless you opt out. Article 7 of China's 2017 National Intelligence Law requires organisations to cooperate with state intelligence work; scholars disagree on how far that compels companies. In April 2025 Korea's privacy regulator found DeepSeek had sent user data, including prompts, to a ByteDance-affiliated company in China without separate consent. In January 2025 security firm Wiz found an unsecured DeepSeek database with over a million log lines, including chat history; DeepSeek secured it promptly once told.
Open weights on your own GPUs. The weights are a file of numbers: prompts and answers stay on your servers unless the software you run them with connects out (see the checklist below). See bare metal vs cloud GPUs.
Who has restricted them
Who | When | What it covers | Restricts private self-hosting? |
|---|---|---|---|
Italy's data protection authority | 30 Jan 2025 | Urgent limit on processing Italians' data | No |
India's Finance Ministry | 29 Jan 2025 | Staff told to avoid AI tools, DeepSeek included, on office devices | No |
Taiwan's digital ministry | 31 Jan 2025 | Agencies, public schools, state firms, critical infrastructure | No |
Australian government (security direction 001-2025) | 4 Feb 2025 | DeepSeek on government systems and devices | No |
South Korea's privacy regulator (PIPC) | 15 Feb to 28 Apr 2025 | New app downloads paused until privacy fixes | No |
Czech government (NUKIB advice) | 10 Jul 2025 | DeepSeek products and APIs in public administration | No; local models with no link to DeepSeek exempt |
US states (Texas, New York, Virginia and others) | From Jan 2025 | State government use; Texas's list (updated Feb 2026) also names Moonshot, Z.ai, MiniMax, StepFun, Baichuan, Alibaba and Baidu | No |
US Defense Department and intelligence community | 2026 defense act (NDAA), 18 Dec 2025; DoD order 8 Sep 2026 | DoD: covered AI from DeepSeek, High Flyer and covered nations, China included; intelligence community: DeepSeek | Only for those bodies and DoD contract work; waivers case by case |
US federal civilian agencies (DHS order 2026-01) | Published 20 Oct 2026 | DeepSeek products, including others' products built with it; binds contractors | Agencies and contractors: its text has no self-hosting exception. Others: no |
The Congressional Research Service says the federal orders limit federal procurement but do not restrict non-governmental use (27 October 2026). Wider bills, such as the No Adversarial AI Act (June 2025), had not passed by November 2026. China's foreign ministry said on 20 October 2026 that it opposes over-generalising national security. Government contractors and regulated firms should take legal advice.
What the licences let you do
The licence table in our GPU sizing guide covers DeepSeek (MIT, no extra terms) and three where only model-as-a-service (MaaS) businesses, which sell access to a model, usually through an API, need sign-off: GLM-5.3 (a Z.ai security review above $10B revenue), Kimi K3.1 Pro (a separate agreement above $20M) and Qwen3.8 2.4T (a separate licence above $50M, for AI coding and office-assistant businesses too). Others:
- Qwen3.8-Flash-Next: any MaaS, AI coding or office-assistant business needs a separate licence, whatever its revenue; internal use is exempt.
- ERNIE-5.5-Flash-Base: no rights in the EU, the UK or South Korea; elsewhere free up to $50M a year of related revenue, with "Powered by ERNIE" shown and a one-time notice to Baidu. No military use.
- MiniMax M3 and M3.1 Flash: commercial use needs a "Built with MiniMax M3" (or "M3.1-Flash") notice and a one-time notice to MiniMax, or approval above $20M a year of product revenue. No military use.
- Step 5 Preview: device makers, and MaaS companies above $20M of combined revenue, need a separate licence.
- Hy4 flash: a separate licence for MaaS operators above $50M of worldwide revenue, or products above 100M monthly users.
- Older DeepSeek releases (2023 licence) bar military use; V4 is MIT.
Revenue tests usually count your whole group, and terms change between checkpoints. This is general information, not legal advice: read the LICENSE file of the exact model.
How good they are
Artificial Analysis's Intelligence Index combines many independent tests into one score. Version 5.0.1, the current one:
Model | Maker | Weights | Score |
|---|---|---|---|
Claude Fable 5.2 (leader, for comparison) | Anthropic (US) | Closed | 65 |
Kimi K3.1 Pro | Moonshot AI | Open | 46 |
Qwen4-Max | Alibaba | Closed | 43 |
Hy4 | Tencent | Not yet released | 35 |
DeepSeek V4.1 Flash (1017) | DeepSeek | Open | 33 |
GLM-5.3 | Z.ai | Open | 32 |
MiMo-V2.6-Pro; Ling 3.1 Flash | Xiaomi; Ant | Open | 30 |
Qwen3.8-Flash-Next; MiniMax M3.1 Flash; Step 5 Preview | Alibaba; MiniMax; StepFun | Open | 29 |
Version 5 launched on 27 October 2026; scores are not comparable across versions. Artificial Analysis puts the cost per task at $26.54 for Fable 5.2, $3.68 for Kimi K3.1 Pro and $0.60 for DeepSeek V4.1 Flash. So Kimi gets about 70% of the leader's score at a seventh of the cost, and DeepSeek about half at a forty-fourth. For the closed US leaders and America's own open models, see US AI models.
What they cost through the labs' own APIs
The labs' own list prices unless marked, in US dollars per million tokens, as of 19 November 2026:
Model | Input | Output |
|---|---|---|
Claude Fable 5.2 (Anthropic, for comparison) | $8 | $40 |
Kimi K3.1 Pro (batch jobs: 60% of list) | $2.00 | $10.00 |
Qwen4-Max, closed (launch price as reported) | $1.50 | $4.50 |
GLM-5.3 (repeated input, cached: $0.26) | $1.40 | $4.40 |
DeepSeek V4-Pro-0813, off-peak / peak | $0.66 / $1.32 | $1.98 / $3.96 |
Step 5 Preview (as listed on OpenRouter) | $1.00 | $2.70 |
MiMo-V2.6-Pro (as listed on OpenRouter) | $0.435 | $0.87 |
MiniMax M3 (up to 512K input tokens) | $0.30 | $1.20 |
DeepSeek V4.1-Flash-1017, off-peak / peak | $0.15 / $0.30 | $0.60 / $1.20 |
MiniMax M3.1 Flash | $0.15 | $0.60 |
GLM-5.3-Flash | $0.15 | $0.50 |
DeepSeek's peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, except Chinese public holidays. Other hosts' prices on OpenRouter vary widely: Kimi K3.1 Pro's input runs from $0.24 to $4.40 across 27 listings (median $2.00), at different precisions and speeds.
APIs this cheap are hard to beat for light use. Running your own pays for steady, heavy traffic, data that must stay on hardware you control, a fixed model version, or fine-tuning. DeepSeek V4.1 Flash and GLM-5.3 fit on one 8-GPU H200 server; Kimi K3.1 Pro and Qwen3.8 2.4T (at 4-bit, a compressed format) need a full 8-GPU B300 server or two H200 servers. Our GPU sizing guide has a model-by-model GPU table; then see rent, buy or API, GPU cost per hour and token pricing for the maths.
What security testing found
US government testing by NIST's Center for AI Standards and Innovation (CAISI) found:
- DeepSeek (September 2025): AI agents (software that acts on its own) built on its most secure model were 12 times more likely than US models to be hijacked by hidden instructions. With a common jailbreak (a prompt trick that gets round safety rules) it answered 94% of overtly malicious requests (US models: 8%). Its models echoed four times as many inaccurate or misleading Chinese Communist Party (CCP) narratives.
- DeepSeek V4 Pro (May 2026): the most capable Chinese model tested, about eight months behind the US frontier.
- GLM-5.2 (July 2026): never hijacked by public attacks; more robust than other open-weight models tested, less than US closed ones; safeguards removable when self-hosted.
- Kimi K3 (with the UK AI Security Institute, July 2026): safeguards did not stop it trying to write exploits (code that attacks software flaws), though it never reached arbitrary code execution (taking over the target) in 41 attempts.
- GLM-5.3 (September 2026): the most capable open-weight model released to date at cyber tasks, such as finding software flaws, but well below current US frontier models (about four months behind).
Security firm CrowdStrike (November 2025) found DeepSeek-R1 up to 50% more likely to write severely vulnerable code when prompts mentioned CCP-sensitive topics, without claiming sabotage. Consultancy Booz Allen (June 2026, via Help Net Security) found Qwen3-Coder wrote about 130% more vulnerabilities when told the user was the US government; Kimi K2.5 scored best, and Booz Allen said its evidence stops short of backdoors.
No state-planted backdoor has been publicly verified in a released Chinese model. But researchers have shown that backdoors can be planted in any open weights, even ones that switch on only once the model is quantised, or compressed (ETH Zurich, 2024).
Censorship and slanted answers
China's 2023 generative AI rules (Article 4) require content to uphold "core socialist values". A February 2026 study in PNAS Nexus found Chinese models refused more questions on Chinese politics and gave shorter, less accurate answers. DeepSeek-R1's censorship persists when self-hosted (R1dacted, May 2025), and CAISI found Kimi K2 Thinking heavily censored in Chinese, much less in English. Censorship can be reduced. Perplexity's R1 1776 (February 2025) retrained it out, with no loss on maths and reasoning by Perplexity's own tests. In one July 2026 test, CTGT found it did not carry over to a smaller model trained on DeepSeek V4 Flash's answers. Re-test after any change.
A checklist for running them safely
Based on guidance from the NSA and allied agencies, CISA, Microsoft with NVIDIA, and F5 Labs (March to October 2026), and the tests above:
- Download from the lab's official Hugging Face page and pin the exact checkpoint.
- Check hashes or signatures; log unsigned files as an open risk.
- Use safetensors weight files, which cannot run code, not pickle files, which can; review any code shipped with the model before running it.
- Avoid unofficial quantised copies, or test them as new models.
- Test in isolation and block outbound internet from inference servers.
- Red-team the model (try to make it misbehave) on your own tasks, including coding prompts naming sensitive topics or a government user, and China-sensitive questions.
- Filter inputs and outputs, keep logs, and have people review generated code and agent actions.
- On an API, set data controls (on API aggregators, choose providers with zero data retention) and keep sensitive data off services you do not control.
How widely they are used
Measure | Figure | When |
|---|---|---|
OpenRouter's 10 most-used models | 8 Chinese (DeepSeek, Tencent, Z.ai, StepFun, Xiaomi, MiniMax) | Week to 18 Nov 2026 |
Hugging Face downloads | 41% Chinese, the largest share | Year before March 2026 |
None of these is market share; see the most popular AI models for the wider picture.
Why they give the weights away
DeepSeek's founder, Liang Wenfeng, said in 2024 that closed-source moats are "temporary" (translated by ChinaTalk). Analysts add US limits on chips, developer goodwill, later cloud and API revenue, prestige and state backing (Scientific American, MIT Technology Review, Takshashila).
Self-hosting DeepSeek, Qwen or Kimi
Chinese open weights keep your prompts in-house only on hardware you control, with outbound internet blocked. We do not host models, sell tokens or run hardware; we find the server. Each is a whole 8-GPU H100, H200, B200 or B300 machine, or a GB300 NVL72 rack on request, from a verified operator in one of our six regions: North America, the UK, Europe, India, Asia Pacific and the Gulf, on a term of at least a year, billed upfront. Tell us which Chinese model and how much traffic, and we will size the server and get quotes, or see our GPU servers.
Where these figures come from
Checked on 19 November 2026 unless dated.
- Models and licences: DeepSeek V4-Pro, V4.1-Flash, Qwen3.8 2.4T, Flash-Next, Kimi, GLM, Hy4 flash, Step 5, MiMo, ERNIE, Seed-OSS, LongCat, DeepSeek 2023 licence; The Next Web (16 Nov 2026); TechNode (4 Nov 2026).
- Prices: DeepSeek, Z.ai, MiniMax, Kimi, Anthropic; others and listing ranges on OpenRouter.
- Scores: Artificial Analysis, cost per task.
- Data: DeepSeek privacy policy, National Intelligence Law (China Law Translate), PIPC (24 Apr 2025), Wiz (Jan 2025).
- Restrictions: Garante, Dawn, Taipei Times (2 Feb 2025), PSPF, NUKIB, Texas DIR, CRS on the NDAA and federal orders, Haynes Boone, CNBC (20 Oct 2026).
- Testing: NIST CAISI on DeepSeek, Kimi K2 Thinking, V4 Pro, GLM-5.2, Kimi K3, GLM-5.3; CrowdStrike; Booz Allen; ETH Zurich.
- Censorship: Interim Measures (China Law Translate), Pan and Xu, R1dacted, R1 1776, CTGT.
- Guidance: Microsoft and NVIDIA, NSA, CISA, F5 Labs.
- Usage: OpenRouter rankings (week to 18 Nov 2026); Hugging Face spring 2026 report.
- Why open: ChinaTalk, Scientific American, MIT Technology Review (21 Apr 2026), Takshashila (7 Aug 2026).